Logan Kilpatrick

like two months ago now at this point from like a training perspective so within that last like month and a half two months whatever it was month um there's been like post-training innovation and a bunch of that like new recipe and stuff like that made it into the flash model that didn't make it into the pro model so a bunch of that stuff will go into the next version of the pro model but it just like from a time sequencing standpoint like didn't end up making it in so there's no like uh

Obviously, the distillation worked really well from Pro to Flash as well, but you'll see those gains also trickle back to Pro, which is great.

Good question.

I mean, I think some of this is, you see this on the whole stack.

Like when you distilled, there's like been a bunch of innovation in distillation when you like take large models and sort of extract a bunch of the like core intelligence and put it into small models.

And like historically that was like a more lossy,

process, which is interesting.

And I feel like it's gotten like much better at like actually making sure a bunch of those capabilities transfer over.

Um, so I think that's part of the story is just like innovation.

Um, I think there's definitely like a data curation aspect.

I don't think that's actually unique to the flash models for what it's worth.

I think that's happening like across everywhere, which is like, how do we do more with less data?

How do we find the data that's actually like really useful and amplify it?

How do you, you know, all that, that whole story I think is playing out.

Um.

I also think there's, and I talked to Oriol and Jeff and Noam yesterday, and I think they made the comment, which is 100% true, which is also you can just do more iteration with the flash models just because they're smaller.

So you get more versus you can't do as much research experimentation with Pro because it's big.

And actually, if you look back to when we did our first reasoning model,

to flash to flash was our first reasoning model back in i think december 2024 and it was explicitly a flash model at that time with reasoning capabilities because the team wanted to be able to get a bunch of shots on goal very very quickly and sort of scale up from that perspective so they figured out a bunch of stuff that worked really well brought it over to pro and then sort of have used that now as the sort of baseline for this stuff so um it is nice to have small models

100%.